Académique Documents
Professionnel Documents
Culture Documents
461–484
This article presents an optimization model of the quantifiable aspects of architectural floorplan layout design, and a
companion article presents a method for integrating mathematical optimization and subjective decision making
during conceptual design. The model presented here offers a new approach to floorplan layout optimization that
takes advantage of the efficiency of gradient-based algorithms, where appropriate, and uses evolutionary
algorithms to make discrete decisions and do global search. Automated optimization results are comparable to
other methods in this research area, and the new formulation makes it possible to integrate the power of human
decision-making into the process.
1 INTRODUCTION
Spatial configuration is concerned with finding feasible locations and dimensions for a set of
interrelated objects that meet all design requirements and maximize design quality in terms
of design preferences. Spatial configuration is relevant to all physical design problems, so it
is an important area of inquiry. Research on automation of spatial configuration includes
component packing [11–13], route path planning [18], process and facilities layout, VLSI
design ½16; 17, and architectural layout [3–10]. Architectural layout is particularly interesting
because in addition to common engineering objectives such as cost and performance, archi-
tectural design is especially concerned with aesthetic and usability qualities of a layout,
which are generally more difficult to describe formally. Also, the components in a building
layout (rooms or walls) often do not have pre-defined dimensions, so every component of the
layout is resizable.
Reported attempts to automate the process of layout design started over 35 years ago [3].
Researchers have used several problem representations and solution search techniques to
describe and solve the problem.
One approach to spatial allocation is to define the available space as a set of grid
squares and use an algorithm to allocate each square to a particular room or activity [4–7]
(see Fig. 1). This problem is inherently discrete and multi-modal. Because of the
ISSN 0305-215X print; ISSN 1029-0273 online # 2002 Taylor & Francis Ltd
DOI: 10.1080=0305215021000033735
462 J. J. MICHALEK et al.
2 OPTIMIZATION OF GEOMETRY
The geometric optimization problem is posed as a process of finding the best location and
size of a group of interrelated rectangular units. A new decision model is formulated
where all objectives and constraints are continuous functions, and all design variables
have continuous domains.
A Unit is defined as a rectangular, orthogonal space allocated for a specific architectural
function. Examples of architectural functions include living spaces, storage spaces, facilities,
and accessibility spaces. For simplicity, this representation assumes that all Units can be
represented as rectangles or combinations of orthogonal rectangles. This simple representation
can model a large array of architectural layouts, and more complex shapes could be added to
the model to expand this array. Figure 2 shows a Unit represented as a point in space ðx; yÞ, and
the perpendicular distance from that point to each of the four walls: fN ; S; E; and W g. This
model has more variables than necessary to describe the shape; however, it allows an optimiza-
tion algorithm to change the position of a Unit independently without affecting its size (by
changing x or y), and it can change any of the four wall positions independently (by changing
N ; S; E, or W ). Although this model increases the problem dimensionality, it offers a lot of
flexibility to make the best design moves at each step of the optimization.
Units are grouped into several categories based on their function: Rooms, Boundaries,
Hallways, and Accessways. Rooms are Units used for sustained living activity as determined
by the designer. The differentiation between living space vs. non-living space is important
only in optimization objectives that maximize the amount of space used for living relative
to all other space. A Boundary is a Unit that has other Units constrained inside of it, and
it is not considered living space. A Hallway is a Unit with no physical walls that is not a
living space. Hallways function as pathways. An Accessway is a Hallway that is constrained
to geometrically intersect two Units. Accessways are generally restricted to be small,
and they are forced to intersect two other Units. They function to keep the two Units adjacent
and connected, and to ensure that there is room for a door or opening between the rooms.
In Figure 3, the external rectangle represents the building Boundary, the living room, bed-
room, and bathroom are Rooms, the hall is a Hallway, and the three Units labeled ‘‘A’’ are
Accessways that define space for a doorway between Units.
Units that are along external walls may also have windows for natural lighting. Window
height can be fixed for each Unit, and window width is a variable. oN ; oS ; oE ; oW represent
the width of the north, south, east and west windows, respectively.
where x is the vector of design variables, n is the number of variables, and h(x) and g(x) are
vectors of equality and inequality constraints.
The window variables drop out when the window is not physically present for a specific Unit
and direction. In order to simplify calculations and notation, several ‘‘intermediate’’ variables
are used to describe geometry that results from the design variables. The following resultant
variables are calculated from the design variables.
These relations are linear, so linear functions of these intermediate variables are also linear
functions of the original variables.
yNi yNj ; ySj ySi ; xEi xEj ; and xWj xWi ð9Þ
The Prohibit Intersection Constraint functions to prevent two Units from occupying the
same space. By default, one Prohibit Intersection Constraint is added for each combination
of Rooms, Hallways, and Accessways, except where two Units are forced to intersect, or
where one Unit is forced inside of another. In order to prevent Unit i from intersecting
Unit j, at least one of the following constraints must be satisfied
The logical disjunction can be represented in negative null form using a min function
ySi yNj ; ySj yNi ; xWi xEj ; and xWj xEi ð12Þ
Although these constraints ensure intersection of the two Units, they permit intersection at a
point. Designers of architectural spaces are generally interested in intersection that provides
enough room for a doorway or opening. To model this, an additional constraint is included to
enforce overlap in one of the Cartesian directions that is at least as large as the doorway or
opening. Therefore, in addition to intersection, at least one of the following conditions must
be satisfied
where di is the minimum size for a door or opening in Unit i. This disjunctive set of
constraints can be represented in negative null form using a min function similar to Eq. (11).
Although this constraint function is nonlinear and non-smooth in part of the design space, it
is linear in most of the design space (similar to Eq. (11)). The complete Force Minimum
Intersection Constraint Group is represented as a set of constraints that force intersection
(Eq. (12)) and another constraint to ensure that the overlap is large enough for access
(Eq. (17)).
The Force To Edge Constraints are used to force a Unit to the edge of a Boundary because
of a window or external door. It is assumed that the first Unit i has already been forced inside
Unit j by another constraint. In order to force a Unit to a particular wall, one of the following
constraints can be added as appropriate.
If connection to an edge is important, but the specific edge is not important, (for instance, a
building may require an external door, but it is not important which direction the door faces),
then the following constraint can be added to represent the disjunction in Eq. (18).
min ðxEi xEj Þ2 ; ðxWi xWj Þ2 ; ðySi ySj Þ2 ; ðyNi yNj Þ2 ¼ 0 ð19Þ
The Minimum Ratio Constraint Group can be used to maintain a desired aesthetic scheme
or prevent long, narrow Rooms that may not be usable. The Minimum Ratio Constraint
Group consists of two constraints.
The Build Cost Constraint is used to keep the construction cost below some value, Gbudget .
For simplicity, build cost is measured only in terms of material cost. Material costs for walls
kwall and for windows ko are specified as dollars per square foot of material, and other costs
are ignored. The build cost constraint is calculated as
where AN ’ AS ’ AE’ AW are the areas of the external walls in each compass direction and
Ao Ao Ao AoW are the areas of windows facing each compass direction. These quantities
N’ S’ E’
are computed in Eq. (30) and Eq. (31).
The Feasible Window Constraints ensure that the window width cannot be larger than the
wall it is on. In addition to the simple bound restricting window size to be positive, this
ensures feasible window size. Each window added to a Unit is given one of the following
Feasible Window Constraints as appropriate.
oNi li ; oSi li ; oEi wi ; or oWi wi ð25Þ
The Bound Lighting Constraint is used to ensure minimum natural lighting for specific
rooms. A simple estimation of the amount of daylight entering a Unit with windows is cal-
culated using environmental and material information. The following procedure is used:
First, available daylight at the window exterior is determined. IESNA [31] provides three
standard skies for use in the evaluation of daylight designs. Approximate available daylight
can be determined from these based on altitude and azimuth angles.
for month m. The coefficient of utilization, CU , a function of the room geometry and window
size, determines the fraction of the available daylight that enters the room. CU can be found
in pre-tabulated data [30] based on room depth, window width, and window height. The net
transmittance for a window facing direction j is calculated as
0:9mG Aoj
mj ¼ ð26Þ
Aj
for room i, where j takes on each direction fN ; S; E; W g, and m spans the 12 months of the
year. The illuminance is then converted into watts, Yi :
Ei 103
Yi ¼ ð28Þ
Ai beff
where Ai is the area of room i, and beff is the efficacy of the light source (assumed to be 80).
The required natural lighting per square foot, yreqi , is defined for each Unit by the designer
(default 1 Watt=sq.ft.). Assuming uniform light distribution, total required natural lighting
can be calculated as Ai yreqi . The minimum percentage of required lighting that is provided
by natural light, jmini , can be specified by the designer. The final constraint is written as:
Yi
jmini ð29Þ
Ai yreqi
468 J. J. MICHALEK et al.
Here D takes on the four directions fnorth, south, east, westg, and UnitD refers to Units that
have windows in direction D. The net area of each external wall is
AD ¼ l1 h1 AoD ð31Þ
Here D takes on the directions fnorth, south, east, westg, and 1 indicates Unit 1, which is
assumed to be the building Boundary Unit. The heat loss calculation assumes that all heat
is lost from the external walls and windows (no heat is lost through the roof ). This model
could be changed depending on what type of building is being modeled. The coefficient
of transmittance for the wall, Uwall , and window, Uo , are tabulated based on the materials
used. The annual heat loss is
X
Qheat ¼ DTi ððAN þ AS þ AE þ AW ÞUwall þ ðAoN þ AoS þ AoE þ AoW ÞUo Þ ð32Þ
i
where i is the set of months where heat is used, and DTi is the average internal=external tem-
perature difference for month i. Finally, the cost to maintain temperature is calculated. Gas
heat is assumed, and the cost of gas per cubic foot, kgas and efficiency of the heater in
Watts per cubic foot of gas, Zheater , can be specified. The heating cost objective function is then
kgas Qheat
minimize Gheat ¼ ð33Þ
Zheater
The Minimize Cooling Cost Objective estimates heat gain during hot months. The proce-
dure for calculating cooling loads is more complicated than heating loads because heat due to
solar gain must be taken into account. The procedure works as follows. First, the net area of
windows on each external wall is calculated using Eq. (30), and the net area of each external
wall is calculated using Eq. (31). Next, the solar heat gain through the windows is estimated.
Several parameters are important in calculating solar heat gain. Depending on the orientation
of the windows (N, S, E, or W), the Solar Heat Gain Factor, bshgf , can be found in tables for a
given location [30]. The shading coefficient, bsc , is a property of the glass [30], and the time-
lag factor, btlf , is a tabulated function of glass type and window orientation [30]. The annual
solar heat gain, Qsolar , is calculated as
!
X
Qsolar ¼ bsc ðAoN bshgf N btlf N þ AoS bshgf S btlf S þ AoE bshgf E btlf E þ AoW bshgf W btlf W Þ ð34Þ
i
ARCHITECTURAL DESIGN 469
where i is the set of months where air conditioning is used. Next, the conductive heat gain
through the building exterior is estimated. The orientation of each exterior wall and windows
is accounted for in the factor. The cooling load due to conduction is calculated as
X
Qcond ¼ DTi ðUo ðAoN btlf N þ AoS btlf S þ AoE btlf E þ AoW btlf W Þ
i
þ Uwall ðAN btlf N þ AS btlf S þ AE btlf E þ AW btlf W ÞÞ ð35Þ
where i is the set of months where air conditioning is used. Finally, the cost to maintain room
temperature is calculated. Electric cooling is assumed, and the rate of electricity, kelec , and
efficiency of the air conditioning unit, Zac , can be specified. The cooling cost objective
function is then
kelec ðQsolar þ Qcond Þ
minimize Gcool ¼
Zac
The Minimize Lighting Cost Objective minimizes the cost spent on lighting the building by
encouraging natural lighting. The amount of natural lighting in room i, Yi , is calculated as in
Eq. (28). The minimum daylight requirement per square foot, blight , is set by the designer
based on usage intention. The total required cost if all of this light is provided by electric
lighting can be calculated as:
!
X
Gelec blighti Ai bH 103 ð36Þ
i
where i is the set of Units, and bH is the number of hours of use per month. The total cost is
then the maximum possible electricity cost minus the cost savings from natural lighting:
!
X
minimize Glight ¼ Gelec Yi bA 103 ð37Þ
i
where i is the set of Units, and bA is the number of hours of available light per month.
The Minimize Wasted Space Objective minimizes building space that is not living space.
This could be space used for hallways or un-allocated space inside the building Boundary.
Wasted space is calculated as the area of the building Boundary minus the total area used
as living space. The objective is formulated as
!
X
minimize l1 w1 li wi ð38Þ
i 2 Rooms
The Minimize Hallway Objective is used to provide extra living space where possible. The
objective is formulated as
!
X
minimize li wi ð40Þ
i 2 Hallways
Multiple objectives can be selected and combined into a single objective function using a
weighted sum of the individual objective functions.
X
N
f ðxÞ ¼ wj fj ðxÞ ð41Þ
j¼1
where fj ðxÞ is the jth objective function, wj is the weighting (relative importance) of the jth
objective function, and N is the total number of objective functions. Appropriate weights may
be difficult to set for objective functions measured in different units. After obtaining results,
weights can be adjusted to compensate and to guide the design to desired results. The objec-
tives presented here do not compete in most of the design space, except for cost objectives,
which are all measured in dollars. This makes multi-objective optimization much easier. In
practice weights only need to be adjusted to keep the function values in the same order of
magnitude to avoid computational problems.
FIGURE 4 Progression of the CFSQP algorithm optimizing a sample apartment complex building to minimize
annual cost and wasted space; (a) shows the initial layout sketch provided by the designer (accessways shown as lines
between Units); (b) is an intermediate feasible iteration (accessways shown also as rectangles); and (c) shows the
completed design (accessways shown as wall openings for clarity).
472 J. J. MICHALEK et al.
FIGURE 4 (Continued)
nature of the formulation, neither algorithm was successful at finding feasible designs for
small problems. This does not mean that the algorithms cannot be successful at laying out
architectural spaces; however, both algorithm are ill-suited to the formulation presented here.
A hybrid SA=SQP search strategy was developed to take advantage of the global qualities
of SA and the efficiency of SQP in order to generate local optima of global quality. The
method works as shown in Figure 5.
In this method, SA is used to search for a good starting point, and SQP is used to find the
local minimum near each starting point. In this way SA can search the space more globally
with large moves while SQP worries about the details. A sample objective function is shown
in Figure 6. In this example, SQP can find six different local optima depending on where the
starting point is chosen. Each point that SA selects is evaluated by locally optimizing it, so
SA observes any point in the vicinity of a local optimum to have the objective value of that
local optimum. In a sense, the objective function is being screened for SA. Notice in the
example that the function SA observes has only two local optima instead of six. Also, an
algorithm searching the resultant function can make larger design moves without as much
danger of overstepping important features.
The hybrid SA=SQP method generated local optima of reasonable global quality for up
to seven room apartment layouts (70 variables, 269 constraints – see Ref. [1] for resulting
layouts). It is important to understand that these designs were generated automatically with
no feasible initial starting point. This is a substantial improvement. Using SA alone, we
were unable to produce even a feasible design. SQP is quick at generating solutions; how-
ever, the designer must define where Rooms should be placed relative to one another.
In this problem, the arrangement is not specified by the designer. The algorithm is
able to automatically generate a quality feasible arrangement and optimize that geometry
locally.
Another way to search for solutions of global quality is to use a variation of an optimiza-
tion technique referred to as the Maximum Distance Distribution Method (MDDM ½23; 24).
This method was developed for discrete problems, but it also works for continuous problems.
The concept is to use a local optimization algorithm to find a local minimum x using the
formulation in Eq. (1). Once the local minimum is found, a new optimization problem is for-
mulated to maximize the distance from x subject to an extra constraint that the new point
must have an objective value at least as good as f ðx ).
maximize ðx x Þ2
subject to hðxÞ ¼ 0;
gðxÞ 0; ð42Þ
f ðxÞ f ðx Þ 0
x 2 <n
If optimizing Eq. (42) yields a solution xy in a new area of the design space, then optimizing
Eq. (1) again with xy as a starting point will often yield a better local minimum. This process
can be repeated by iteratively solving Eqs. (1) and (42) to obtain better solutions. MDDM is
not guaranteed to converge to the global optimum; however, in practice there are many
situations where this method is successful at improving the quality of the local optimum
returned. An example is provided in Figure 7. The method is especially useful if f ðx Þ is
flat in some feasible direction at x .
474 J. J. MICHALEK et al.
FIGURE 7 Demonstration of the MDDM method for finding improved local optima. An initial design (a) was
optimized using CFSQP. The result is a local optimum (b) (the design cannot be improved by small changes in the
design variables). The MDDM method was used to generate design (c), an improved local optimum for this example.
Another design exploration program was written to produce design alternatives by searching
the space using a strategy of random design changes. The program makes design moves of
three types: (1) swap the positions of two Units, (2) perturb the position of a Unit, and (3)
reduce the size of a Unit. After each design move, the program attempts to re-optimize
using the geometric optimization algorithm. The algorithm first attempts to find a feasible
design using penalty methods. If it is unable to find a feasible design, the program makes
one of the three design moves at random. When a feasible design is found, it is saved, and
the program continues by making more random design moves. This strategy was used to
generate designs for a simple three-bedroom apartment layout. The program generated 200
design alternatives overnight. Although this strategy is not rigorous, it is a useful tool for
generating a spread of design alternatives that can be explored further with the interactive
design tool (see Ref. [2]).
3 OPTIMIZATION OF TOPOLOGY
The topology optimization problem is presented as a process of finding the best set of rela-
tionships between rooms in a space. In this formulation, relationships include connectivity,
and initial rough location. Connectivity defines which rooms are directly connected by a
doorway or open pathway. Rough location defines rough arrangement of rooms. Other mod-
els (½9; 10) have used decision variables to define topological spatial relationships (i.e. adj-
to-north-of, adj-to-south-of, etc:). However, the use of rough room position to describe spa-
tial relationships does not enforce these relationships during geometric optimization, so the
geometric optimization algorithm has more freedom to manipulate the geometry.
Topologies could be evaluated based on topological qualities, such as openness, proximity,
directionality, or symmetry; however, even though these aspects are often thought of as
topological, they are difficult to evaluate without rough geometry. It is best to evaluate
objectives using a geometric layout, therefore each topology is evaluated based on the
ARCHITECTURAL DESIGN 475
best geometry that can be generated from it. Using this method, layouts can be optimized for
any objective that can be formulated in terms of geometry or topology.
Figure 8 shows the topology optimization process. A discrete optimization algorithm uses
information from previous topologies to generate new topologies. Each new feasible topo-
logy, X, is translated into a geometric optimization problem. A locally optimal geometry
x is found using CFSQP, and the quality of that geometry fg ðx Þ defines the quality
of the topology that generated it, ft ðX Þ. The discrete optimization algorithm searches for
the topology that generates the best geometry.
xi ; yi 2 Zþ
fij 2 f0; 1g ð43Þ
8 i 2 froomsg; 8 j 2 ðfrooms > ig [ fextwallsgÞ
where ðxi ; yi Þ represents integer Cartesian coordinates of room i, and fij represents the
existence of a connection between room i and room j (or external wall j). Figure 9
FIGURE 9 A 4-room example showing design variables in the topology formulation. (a) Room position grid
showing (x, y) for each room. Phantom lines show room connections. The dashed line shows the implied boundary.
shows a visual representation of the design variables. It is important to note that topological
decisions about relative positions of rooms (i.e., room i is-north-of room j) are represented
here using absolute positions of rooms. Several other methods of representing topological
decisions (½9; 10) do not use absolute positions; however, it is necessary in this strategy
because the geometric optimization algorithm requires a starting design with geometric
information. The use of absolute positions has several consequences: (1) The mapping
from topology to geometry is not injective (one-to-one). It is possible for more than
one topology to generate the same geometry. This means that computation time can be
wasted searching similar topologies. (2) The mapping from topology to geometry is not
surjective (onto). Because each room is represented as a grid point, each topology could
be interpreted geometrically in several ways. It is not clear, however, that every possible
ARCHITECTURAL DESIGN 477
geometric alternative can be generated using the topology definition in this paper. (3) The
space of topology combinations is exponential.
Because of these limitations, this representation is not well suited to problems where all
solutions need to be enumerated. It is not clear that the representation can enumerate all pos-
sible topology alternatives; however, this method is powerful for larger problems where heur-
istic search is necessary. This is because in practice heuristic search algorithms can often find
reasonable quality designs quickly, while enumeration algorithms must systematically
explore designs one by one, which can often take too long to be practical.
Connectivity Constraints are defined by the designer for each problem. The constraints
describe how a certain room is required to be connected to an outside wall, to another
room, or how certain rooms are required to not be connected. For example,
fiN þ fiS þ fiE þ fiW 1 room i required to connect to at least one external wall ð47Þ
Path Constraints are defined by the designer for each problem. A path may be required
between all combinations of rooms, or a path may be required between certain rooms. For
example, a path could be required from the bedroom to the kitchen without passing through
a bathroom or closet. These constraints involve room connectivity, and they are generated for
each specific constraint with an algorithm (see Ref. [1]).
Planarity Constraints ensure that the geometry can be realized with a two-dimensional
(planar) floorplan. One way to ensure planar feasibility is to draw lines between connected
nodes on the position grid and ensure that no two lines cross. These lines will be allowed
to share endpoints as long as they do not share any interior point. This constraint is difficult
to represent with a closed form mathematical function. (see Ref. [1])
Envelope Constraints ensure that Units that are forced to be connected to an external wall
must lie on the external envelope of Units on that wall. The four constraints below are added
for each unit i
3.1.3 Objective
The objective of the topology optimization problem is to minimize the objective value of the
resultant local optimal geometry formed by the topology.
where fg is the objective value of the geometry, and SQP is a function that returns the local
optimum geometry for the topology X. Notice that x and y determine starting locations for
rooms in the geometry formulation while f defines constraints for the geometry as well as win-
dows and accessways (see Fig. 10). The optimization objective can be anything defined by the
geometry. Typically, this research has optimized for the topology that produces the most cost
efficient layout. Only feasible topologies are passed to the geometric optimizer (SQP). If the
topology violates any constraints, then the design is evaluated using penalty functions.
FIGURE 10 Schematic showing the relationship between the topology and geometry optimization algorithms.
ARCHITECTURAL DESIGN 479
topologies return positive objective function values, and feasible topologies that result in bet-
ter geometries (lower fg ðxÞ) have a better objective function value (higher ft ðX Þ).
4 DEMONSTRATION EXAMPLE
A realistic problem was implemented to test the scalability of the automated building genera-
tion algorithm. The example involves a small apartment complex with three separate apart-
ments. Rooms and specifications are shown in Table I. Constraints that are specific to this
problem are listed in Table II. This problem was run for 20,000 generations (100 designs
each generation) to search for global solutions. Feasible designs take much longer to evaluate
than infeasible designs (because feasible designs are passed to the geometric optimization
algorithm), so a second termination criterion was added to terminate after 50 feasible designs
were found. This criterion was intended to make search time more consistent between runs.
The sample topology and resulting geometry solution shown in Figure 12 were generated
using the automated design tool.
The algorithm generates local optimal solutions, but the global search is quite limited due
to combinatorial complexity. Once a feasible topology is found, it has a much higher prob-
ability of being selected as a parent design by the evolutionary algorithm because it has a
much higher fitness value than infeasible designs. Thus, new designs tend to be very similar
to the first feasible design found, and other designs are usually discarded. The result is that
the algorithm tends to fixate on the first feasible solution it finds, exploring mostly variations
of that solution. The algorithm can be run several times to produce design alternatives, but
generally when it is run once, the final population converges to variations of one main design
theme. This is a serious limitation for global search, and the algorithm is more useful as a
feasible-design-finder than as a true optimizer. For smaller problems, the evolutionary algo-
rithm is still able to search a significant range of the design space to find global quality solu-
tions. For this problem, the evolutionary algorithm can consistently find solutions in under
20,000 generations (2 106 design evaluations).
The geometric optimization problem does not have the same combinatorial nature that the
topology problem has, and it is able to handle much larger problems. The example shown
in Figure 4, contains 23 rooms, three hallways, one boundary, and 25 accessways for a
total of 52 units. This geometric optimization problem contains 312 variables and 1578 con-
straints.
FIGURE 12 Sample design topology and final geometry generated by the automated design tool.
482 J. J. MICHALEK et al.
FIGURE 12 (Continued)
5 CONCLUSIONS
Two automated optimization algorithms have been used to automate the generation of design
layouts: the geometry and topology algorithms. The geometry algorithm, built on rigorous
gradient-based algorithms, is efficient and robust, and has been successful at optimizing geo-
metry for large problems. In its present state, it is most useful as an aid for design explora-
tion, rather than design automation, because results are highly dependent on the starting point
defined by the designer. Several tools have been implemented for searching the geometric
space more globally, including a hybrid SA=SQP algorithm and a strategic programming
method. These tools have been successful at automatically finding alternative arrangements
for rooms and exploring many local minima.
A second topology optimization algorithm was built on top of the geometric algorithm to
search feasible topology alternatives and find the feasible topology that generates the best
geometry. The results are interesting, but limited. One advantage to the approach presented
here is that the final design generated by the algorithm can be used as a starting point for
interactive design exploration (see Ref. [2]).
Possible improvements include the addition of new shapes, objectives, and constraints to
the design toolbox to address more complex geometry, material selection, building codes,
structural elements, routing of wires, pipes and ducts, and more accurate models.
Additionally, the topology search can be improved if the topology can be defined in a new
ARCHITECTURAL DESIGN 483
way so that topology decisions create linear constraints in the geometry optimization pro-
blem. Using a different definition of topology, such as in Refs. [3–5], it is possible to elim-
inate the rough position topology variables, and ensure that the mapping from topology to
local optimal geometry is both injective and surjective (meaning that each valid topology
will create a different locally optimal geometries, and that all possible local optimal geometry
could be created by a specific topology).
Acknowledgements
Special thanks to Professor Kazuhiro Saitou, John Whitehead, Panayiotis Georgiopoulos, and
Adam Cooper for their contributions to the topology optimization model and solution
strategy.
References
[1] Michalek, J. (2001). Interactive layout design optimization. MS thesis, University of Michigan.
[2] Michalek, J. and Papalambros, P. Y. (2002). Interactive design optimization of architectural layouts. Engineering
Optimization (this issue).
[3] Levin, P. H. (1964). Use of graphs to decide the optimum layout of buildings. Architect, 14, 809–815.
[4] Liggett, R. S. and Mitchell, W. J. (1981). Optimal space planning in practice. Computer-Aided Design, 13(5),
277–288.
[5] Sharpe, R., Marksjo, B. S., Mitchell, J. R. and Crawford, J. R. (1985). An interactive model for the layout of
buildings. Applied Mathematical Modeling, 9, 207–214.
[6] Jo, J. H. and Gero, J. S. (1998). Space layout planning using an evolutionary approach. Artificial Intelligence
in Engineering, 12, 149–162.
[7] Jagielski, R. and Gero, J. S. (1997). A genetic programming approach to the space layout planning problem.
CAAD Futures, 875–884.
[8] Baykan, C. A. and Fox, M. S. (1997). Spatial synthesis by disjunctive constraint satisfaction. Artificial
Intelligence in Engineering Design, 11, 245–262.
[9] Schwarz, A., Berry, D. M. and Shaviv, E. (1994). Representing and solving the automated building design
problem. Computer-Aided Design, 26(9), 689–698.
[10] Medjdoub, B. and Yannou, B. (1999). Separating topology and geometry in space planning. Computer-Aided
Design, 32, 39–61.
[11] Yin, S. and Cagan, J. (2000). An extended pattern search algorithm for three-dimensional component layout.
Transactions of the ASME, 122, 102–108.
[12] Cagan, J., Degentesh, D. and Yin, S. (1998). A simulated annealing-based algorithm using hierarchical models
for general three-dimensional component layout. Computer-Aided Design, 30(10), 781–790.
[13] Szykman, S. and Cagan, J. (1997). Constrained three-dimensional component layout using simulated annealing.
ASME Transactions, 119, 28–35.
[14] Choo, H. J. and Tommelein, I. D. (1999). Space scheduling using flow analysis. Proc. Seventh Annual
Conference of the International Group for Lean Construction, 299–312.
[15] Rong, B. (1997). Automated generation of fixture configuration design. Journal of Manufacturing Science and
Engineering, 119, 208–219.
[16] Koide, T. and Wakabayashi, S. (1999). A timing-driven floorplanning algorithm with the Elmore delay model for
building block layout. Integration, the VLSI Journal, 27, 57–76.
[17] Wang, Y. and Wu, H. (1998). Method of constraint graphs used in spatial layout. Ruan Jian Xue Bao Journal
of \Software, 9(3), 200–205.
[18] Ito, T. (1999). A genetic algorithm approach to piping route path planning. Journal of Intelligent Manufacturing,
10, 103–114.
[19] Fadel, G. M., Sinha, A. and McKee, T. (2001). Packing optimisation using a rubberband analogy. Proceedings of
DETC’01 2001, ASME Design Engineering Technical Conferences, DETC2001=DAC-21051.
[20] Kim, J. J. and Gossard, D. C. (1991). Reasoning on the location of components for assembly packaging. Journal
of Mechanical Design, Transactions of the ASME, 113(4), 402–407.
[21] Chapman, C. D., Saitou, K. and Jakiela, M. J. (1994). Genetic algorithms as an approach to configuration and
topology Design. Journal of Mechanical Design, Transactions of the ASME, 116(4), 1005–1012.
[22] Papalambros, P. Y. and Wilde, D. J. (2000). Principles of Optimal Design–Modeling and Computation, 2nd ed.
Cambridge University Press, Cambridge, England.
[23] Kott, G. J. and Gabriele, G. (1998). A tunnel based method for mixed discrete constrained nonlinear
optimization. ASME Design Automation Conference, DETC98=DAC-5592.
[24] Kott, G. J. (1998). A method for mixed variable constrained nonlinear optimization based on a new function
bounding technique. PhD thesis, Rensselaer Polytechnic Institute.
484 J. J. MICHALEK et al.
[25] Zhou, J. L. and Tits, A. L. (1995). An SQP algorithm for finely discretized continuous minimax problems and
other minimax problems with many objective functions. SIAM Journal of Optimization, 6, 461–487.
[26] Lawrence, C. T., Zhou, J. L. and Tits, A. (1995). User’s guide for CFSQP version 2.3: A C code for solving large
scale constrained nonlinear minimax optimization problem, generating iterates satisfying all inequality
constraints. Institute for Systems Research, University of Maryland, Technical Report TR-94-16r1.
[27] Bartak, R. (1999). Constraint programming: In pursuit of the Holy Grail. http:==kti.ms.mff.cuni.cz=
bartak=constraints.
[28] Wall, M. (1996). GAlib: A C þ þ library of genetic algorithm components-Version 2.4 -Document revision B.
galib-request@mit.edu, http:==lancet.mit.edu=ga=dist=galibdoc.pdf.
[29] Bentley, P. J. (1999). Evolutionary Design by Computers. Morgan Kaufmann Publishers, Inc., San Francisco, CA.
[30] ASHRAE (1997). Fundamentals. American Society of Refrigeration, Heating and Air-conditioning Engineers,
Atlanta, GA.
[31] IESNA (1998). IESNA Lighting Education: Intermediate Level. Illumination Engineers Society of North
America.